SAIG CSO Testifies Before Canada's House of Commons on Sustainable AI

Honourable Members of the Committee, thank you for convening this crucial study. 

As an AI researcher based in Montreal, my career has been dedicated to measuring the environmental impacts of Artificial Intelligence – first in academia, working at the Mila institute with Dr. Yoshua Bengio, but now in industry, where I recently co-founded the Sustainable AI Group, a research and advisory company dedicated to helping organizations measure and reduce the environmental impacts of the AI that they use.

In the last decade, my research has increasingly exposed  a critical, often ignored reality: AI is a physical technology. It does not exist in an ethereal cloud; it relies on massive physical infrastructure, consumes large amounts of energy and water, and leaves behind a significant carbon footprint.

As Canada reshapes its digital policy agenda in 2026 following the legislative reset of Bill C-27, we must understand that digital sovereignty and environmental sustainability are two sides of the same coin. Canada is uniquely positioned to lead the global shift toward responsible, sustainable AI. We should not attempt to out-spend foreign monopolies on bloated, general-purpose models and gigawatt datacenters. Instead, our competitive advantage lies in building transparent, green, and clean AI.

1. AI in Strategic Industries: The Right Model for the Right Task

The integration of AI into sectors like construction and manufacturing offers immense productivity gains. However, the current industry trend—relying on massive, generic, cloud-hosted generative models—is both ecologically and operationally unsustainable.

The large language models that many of us use today are trained at great cost to be general-purpose, which definitely makes sense from the point of view of the tech companies that want to respond to any kind of query thrown at them, be it coming up with a chocolate chip cookie recipe or an itinerary for a family vacation in Italy. These companies can and do spend hundreds of millions of dollars on compute in order to develop these models, often using our own data to improve them and then sell them back to us at a premium price. But most of what businesses in Canada want to use AI for isn’t general-purpose at all, it is specific tasks that require robust, trustworthy technology without all the bells and whistles. 

And in fact, querying generic, multi-billion-parameter generative AI models to optimize a manufacturing assembly line or analyze a construction blueprint is like taking a helicopter to do your groceries. It is an absurd waste of energy. My research has found that generating a high-quality image with AI can use as much energy as half of a cell phone charge, and generating videos requires thousands of times more. And the gap between the most and least efficient models is growing with the rise of reasoning models and agentic AI. However, Canada has an opportunity to incentivize on-device AI and small, task-specific models that run locally within Canadian facilities. This creates a direct, powerful synergy between operational security and environmental sustainability.

2. Federal Legislation: Mandating Transparency and Energy Efficiency

Furthermore, we can, and should, mandate transparency to audit the true environmental and operational costs of training and deploying AI models. When we know where models are running, along with details regarding hardware and energy grid metrics, we can accurately measure sustainability.

The difference between training a multi-billion parameter AI model on a low-carbon grid like those in Quebec and Ontario, versus a grid powered by behind the meter natural gas can mean orders of magnitude fewer emissions. Therefore, future federal AI legislation must mandate that companies deploying high-impact AI models disclose their full lifecycle environmental footprint, including the energy grid mix and water usage effectiveness (WUE) of the hosting data centers.

Canada can foster a booming local ecosystem by supporting companies that use open-source tools like CodeCarbon, which I helped develop, to publicly document their models' energy efficiency. This establishes a Green AI stamp of approval, appealing to a global market increasingly desperate for ESG-compliant technology, differentiating Canada's AI offerings and giving our companies a competitive advantage. The federal government can lead the way by requiring AI developers to provide this information when applying for tender offers for government contracts, setting a baseline for the entire field.

Also, just as we rely on Energy Star ratings for appliances, Canada can implement efficiency certifications for commercial AI models, disincentivizing the use of bloated models for simple industrial tasks. This approach is practical and highly achievable; it has already been proposed by the AI Energy Score project, which I have created and co-led for the last years. In our work, we have tested hundreds of open-source AI models across a dozen tasks, finding efficiency differences in the tens of thousands between models of different sizes and architectures. We need to build upon this approach and test proprietary models as well, allowing Canadian users to choose tools with true energy efficiency in mind.

3. Greening the Government's Sovereign Compute Infrastructure Program 

Finally, we must align the government's current industry plan with these ecological limits. The federal government's allocation toward the Sovereign Compute Infrastructure Program, or SCIP, is a vital step toward reclaiming our data sovereignty. However, compute power cannot be decoupled from environmental boundaries.

The government must ensure that any public supercomputing infrastructure built under SCIP is powered, at least in majority, by renewable energy and utilizing cooling systems that sustainably use the local water supply, and developed in consultation with local residents, including Indigenous communities. Instead of copying the hyperscale, football-field-sized data centers that are becoming the norm in the United States, Canada has the opportunity to fund more creative approaches to compute.  For instance, we can build smaller datacenters that are better integrated with existing infrastructure, allowing us to reduce resource consumption while reusing the generated heat for offices, residences, and university campuses.

Furthermore, SCIP resources should explicitly prioritize Canadian researchers and local open-source initiatives developing climate-tech solutions and sustainable industrial applications, rather than subsidizing the computational overhead of foreign commercial interests. Each project that plans to use SCIP compute should be required to measure and report their energy and emissions, improving the transparency of the field while providing vital data points on resource use across different AI modalities.

Conclusion

True digital sovereignty is completely impossible without environmental sustainability. Canada must reject the current trajectory of AI, which is unsustainable from all perspectives, and define a different trajectory for Canadian AI. 

By legally mandating transparency, championing on-device efficiency, and leveraging our clean energy grids responsibly, Canada can set the definitive global standard for the future of sustainable, sovereign industrial technology.

Thank you, and I welcome your questions.